This is an opinion piece.

For a few days in June, the argument for open-source AI models seemed to write itself.

Just days after Anthropic launched Fable 5, the public-facing version of its more powerful Mythos-class models, the US government moved to suspend it due to national security concerns. Anthropic responded by disabling access to both Fable 5 and Mythos 5.

Access to both models has since been restored, but the episode still exposed the problem of depending on closed APIs.

For open-source advocates, this was the clearest illustration yet of their core argument: a single government directive was enough to cut off access for millions of users overnight.

Image credit: Ulla

To be fair, governments restrict foreign digital services all the time. When TikTok faced a US ban, users found alternatives within days.

But social media platforms are interchangeable in a way that frontier AI models are not. A business built around one frontier model can’t simply swap in another. These models are different enough that replacing one can break the product or significantly degrade its performance.

At least one US company experienced this and decided to sue the US government, arguing that the directive caused immediate and severe harm to its business.

Online chatter has pointed to how this benefits Chinese AI labs, who have mostly relied on open-source releases for better distribution but have trailed US firms in terms of model capabilities. The restriction gives Chinese companies a strong narrative centered around accessibility being the most important feature.

Z.ai is a clear example. The Beijing-based lab released its latest model, GLM-5.2, for its paid subscribers just one day after the Fable 5 suspension, leaning heavily into that exact narrative.

Moving forward, it’s tempting to conclude that Chinese AI labs will simply double down on open-weight models. DeepSeek, for example, remains committed to its open-source strategy, The Information reported.

But the situation isn’t that clear cut. The distribution strategies of Chinese AI labs will have less to do with whether US models will become more restrictive. Instead, it will be about how they respond to competition from an unlikely source: their own inference providers.

Victim of its own distribution?

Going open-source has proven to be an effective customer acquisition strategy for Chinese AI firms so far. OpenRouter, a platform that lets developers compare and access AI models from different providers, gives us an indication of this.

Since mid-February, open-weight models have been more popular than closed models on the platform, judging by the number of tokens generated. Our analysis shows that in May alone, the former accounted for 60.8% of all tokens generated on the platform, compared with 39.2% for closed models.

But there’s a flip side to this strategy.

Once a lab publicly releases a model’s weights or parameters, it allows anyone to download and run it independently. Any inference provider with enough GPUs and a sales team can host the same model – often cheaper, faster, and more reliably than the labs themselves.

These providers specialize in running AI models for users and businesses. Think of AI labs as film studios that make the actual product, while inference providers are streaming platforms that distribute the movie to viewers.

Unlike AI labs, inference providers can run active sales operations and go directly after the same customers, according to an executive from a major Chinese AI lab.

“They can reach out to your customer and say, I can give you a 50% discount,” the person says, adding that some inference providers are already undercutting suggested pricing on open-weight models from Chinese labs.

A quick search on OpenRouter also supports this claim. Third-party providers like DeepInfra or Parasail often offer K2.7 Code, the code-focused version of Moonshot AI’s Kimi model, at around 20 cents per million tokens cheaper than the developer, for instance.

These providers can offer a discount because they can lower the cost of inference – the process of running a trained AI model to generate outputs from new inputs – in two main ways.

For one, it can run the models on optimized hardware infrastructure. It can also host unofficial quantized versions, which are compressed variants that need less computing power and memory. That can make them cheaper to run, though sometimes at a small cost to accuracy.

The Chinese AI lab executive mentioned earlier sees this kind of maneuver as a potential trigger for labs to reconsider whether to keep flagship models open-weight.

Price, however, isn’t always the deciding factor.

Hokiman Kurniawan, co-founder and CEO of AI note-taking startup Meeting.ai, says his company uses open-weight models through a provider called Fireworks. It does this because the service is “more reliable” and “less error-prone” than going directly to some model labs, even if it does cost an extra 10 to 20 cents per million tokens.

There’s another reason Kurniawan prefers Fireworks: geography. Because many of his paying customers are in the US, hosting inference outside China helps ease potential concerns about where client data is processed. He also avoids quantized models, believing they’re less capable than the full versions.

A price war over someone else’s model

So if inference providers can serve the same model cheaper and more reliably, what’s left for the labs to sell? One counterargument is that inference was never meant to be the main business for AI firms.

Kyle Qi, an investor at Llama Ventures, argues that the best labs should monetize enterprise deployments and cloud subscriptions, rather than just charging for model usage. Alibaba, for instance, can afford to open-source Qwen if it helps sell the firm’s cloud computing services.

“With open-weight models, whoever has the cheapest and most reliable chips serves the model best, and increasingly, that is a well-capitalized inference provider, not the lab,” Qi says.

That can leave pure-play AI labs in a more vulnerable position. Unlike Alibaba or ByteDance, companies such as MiniMax, Moonshot AI, and to a lesser extent, DeepSeek and Z.ai depend more heavily on API revenue and have fewer alternative businesses to subsidize their open-source models.

Nevertheless, DeepSeek might be the exception, according to Qi.

The renowned Chinese startup recently raised US$7.4 billion, its first outside capital, but almost half of that came from its own CEO’s pocket. Its ability to cut the prices of its flagship model by 75% also shows how DeepSeek is better positioned to face this inference price war than its peers.

Meanwhile, Qi notes that Z.ai shouldn’t be read as a single business. The majority of its revenue, after all, still comes from private deployments to Chinese state-owned enterprises and financial institutions.

“That business is sticky, regulated, and local, and an offshore inference provider cannot touch it,” Qi adds.

See also: The real money in AI isn’t in the models

Still, Z.ai has a lot at stake now that it’s a public company. In mid-June, its market capitalization crossed the HK$1 trillion (US$127.5 billion) threshold for the first time. That figure is very high relative to its revenue, which analysts estimate could reach HK$3.2 billion (US$408 million) this year.

At roughly 300x its projected revenue, its valuation implies investors are betting on growth that’s strong enough to keep demand from shifting to third parties.

If much of that demand is captured by third-party providers instead, the company risks losing part of that upside. That doesn’t mean Z.ai will abandon open-weight releases, but the valuation pressure should give the company more reason to protect the parts of the business it can directly monetize.

The game theory of staying open

A former researcher at a major Chinese AI lab says that the decision to open-source is “never purely a technical one.” There are other factors at play, like the company’s brand, its ecosystem, talent, and monetization risks.

Qi of Llama Ventures argues that one of three things need to happen before a Chinese AI lab veers away from open-source. One, it develops a model so far ahead that giving away the weights becomes too costly. Two, it builds an independent monetization engine that no longer relies on open distribution. Or three, if there’s pressure from the state to do otherwise.

He adds that a version of this is already happening, though it hasn’t exactly been a clean shift from open to closed. He points to models such as Z.ai’s GLM-5-Turbo and Alibaba’s Qwen3.5-Omni as examples of companies keeping some advanced models or hosted versions under tighter control, even as they continue to release other models with open weights.

Another path is to keep things open while tightening the commercial terms. Moonshot’s Kimi K2.5 is a textbook case.

Moonshot released that model under a modified MIT license. This allows for broad free use but requires prominent attribution once a commercial product that is built around it crosses roughly 100 million monthly active users or US$20 million in monthly revenue.

Nevertheless, going closed has its own problems, the executive from a Chinese AI lab mentioned earlier notes. If one lab goes closed-source while its rivals remain open, it can continue learning from the open ecosystem while capturing more of the economics itself.

If every lab makes the same calculation, however, the ecosystem that accelerated China’s AI progress might suffer. “There’s no single causation between your act and your outcome,” the person says. “You’re influenced by a lot of players.”

The suspension of Fable 5 made open weights look stronger strategically. But the more successful that strategy becomes, the harder it is for model labs to capture the value they generate.

What we can be certain of is that Chinese AI labs must overcome this dilemma.

Currency converted from Hong Kong dollar to US dollar: US$1 = HK$7.84.

Which companies in India are blazing a trail for the rest? One way to judge – although imperfect – would be the amount of money they have raised. Using Tech in Asia’s data, we’ve generated this constantly updated list of 50 startups and tech companies in India who have raised the most money from investors.

To keep the list fresh, we’ve only included funding data from the past two years.

Seeking more? Search the most comprehensive database of tech companies in Asia or read our research methodology. For more auto-generated lists of funding data, head here.